Real-time supervision system of intelligent box-type substation

Abstract

The invention relates to the technical field of intelligent monitoring of power equipment, in particular to a real-time monitoring system for an intelligent box-type substation, which comprises a multi-source data acquisition module, a voiceprint fault diagnosis module, an anti-theft space-time prediction module, a collaborative arbitration module and a database. The voiceprint fault diagnosis module removes noise based on an adaptive filter of an LMS (Least Mean Square) algorithm, extracts a Mel-frequency cepstral coefficient and a wavelet packet energy feature, and fuses the Mel-frequency cepstral coefficient and the wavelet packet energy feature to form a voiceprint feature vector; the fault classification unit takes MobileNetV2 as a model and classifies real-time voiceprint features and identifies fault types after training, the anti-theft time-space prediction module creates an LSTM model to output a vibration baseline value and calculates a real-time deviation degree, a vibration propagation path is analyzed by means of a graph convolution network to obtain a space anomaly index, a comprehensive threat value is calculated, and the anti-theft time-space prediction module performs anti-theft and anti-theft on the basis of the comprehensive threat value. The collaborative arbitration module divides rule types according to TAS and voiceprint fault types, fault diagnosis and anti-theft monitoring are effectively achieved through the system, and the safety and reliability of operation of the box-type substation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring of electric power equipment, and in particular to a real-time monitoring system for an intelligent box-type substation. Background Art

[0002] In the power supply system, box-type substations are key facilities, and their stable operation plays a vital role in ensuring reliable power supply. However, the supervision methods of traditional box-type substations have many drawbacks and cannot meet the safety and intelligence requirements of modern power systems. Traditional fault monitoring systems rely on a single type of sensor and a simple threshold judgment mechanism, which means that they can only detect some relatively obvious faults, and are often difficult to detect early and subtle fault signs. These faults gradually develop into serious faults, causing power outages, affecting the continuity of power supply, and bringing inconvenience to industrial production and residents' lives.

[0003] In terms of security protection, traditional systems have weak monitoring and prevention capabilities for theft, often relying solely on simple access control systems and limited video surveillance, and are unable to provide real-time and comprehensive awareness of potential theft threats.

[0004] In summary, it is of great significance to develop a real-time monitoring system for box-type substations that can realize multi-source data fusion, accurate fault diagnosis, effective anti-theft monitoring, and intelligent collaborative decision-making. Summary of the Invention

[0005] In order to solve the technical problems raised by the above background technology, the present invention provides a real-time monitoring system for an intelligent box-type substation.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] The present invention is a real-time monitoring system for an intelligent box-type substation, comprising a multi-source data acquisition module, a voiceprint fault diagnosis module, an anti-theft spatiotemporal prediction module, a collaborative arbitration module and a database.

[0008] The multi-source data acquisition module collects voiceprint data, vibration data, and environmental data in real time based on multimodal sensors. The specific process is as follows:

[0009] A directional MEMS microphone array is installed on the transformer winding side or the circuit breaker operating cavity to collect the original sound wave signals of equipment operation and abnormal sounds in real time; an ultrasonic sensor is installed near the insulator to collect and ultrasonic pulse signals in real time, which can be used to detect microsecond ultrasonic pulse groups caused by insulation defects in the future; three-axis accelerometers are arranged at the four corners of the box or the door axis to collect three-axis vibration values in real time to reflect the intensity and direction of mechanical vibration; a millimeter-wave radar is installed on the rotating pan-tilt head on the top of the box to collect the movement path of the corresponding personnel in real time; the original sound wave signal and ultrasonic pulse signal are integrated into a voiceprint signal, which is sent to the voiceprint fault diagnosis module and voiceprint library and timestamp is added; the three-axis vibration value is sent to the anti-theft spatiotemporal prediction module; the movement path of the personnel is sent to the collaborative arbitration module.

[0010] The voiceprint fault diagnosis module extracts and analyzes the features of the voiceprint data collected during the operation of the box-type substation to accurately identify the type of fault. The specific process is as follows:

[0011] The voiceprint fault diagnosis module is equipped with a pre-processing unit, a feature extraction unit and a fault classification unit;

[0012] The pre-processing unit uses an adaptive filter to separate the noise from the voiceprint signal. The noise includes fan noise and traffic noise outside the substation. The adaptive filter is equipped with an LMS algorithm, whose goal is to minimize the mean square error between the filter output and the expected signal. Specifically:

[0013] Assume the input signal is β(n), the desired signal is γ(n), and the output of the adaptive filter is α(n). The error signal e(n) is obtained using the formula e(n) = γ(n) - α(n). By continuously adjusting the coefficients of the adaptive filter, the mean square value of the error signal is controlled to a minimum, and noise is filtered out from the signal. The filter coefficient update formula is ω(n+1) = ω(n) + ue(n)β(n), where ω(n) represents the filter coefficient and u represents the step size factor.

[0014] The feature extraction unit extracts the Mel-frequency cepstral coefficients and wavelet packet energy features. Specifically, for the Mel-frequency cepstral coefficients, the input voiceprint signal is pre-emphasized through a first-order FIR filter. The formula is K(n) = β(n)-Q(n-1) to obtain the pre-emphasized signal K(n), where Q is the pre-emphasis coefficient with a fixed value of 0.95. The meanings of the other parameters are the same as those in the adaptive filter formula. The purpose of pre-emphasis is to enhance the energy of the high-frequency part, compensate for the high-frequency attenuation of the signal during transmission, and make the signal spectrum flatter; the pre-emphasized signal The signal is divided into several short frame signals of fixed length, and each short frame signal is subjected to fast Fourier transform FFT to obtain the discrete spectrum X(k) of the short frame signal, where k represents the discrete frequency point after FFT transformation, k=0, 1, ..., N-1, and N-1 represents the total number of discrete frequency points. The discrete spectrum is mapped to the Mel frequency scale, and a group of filters are evenly distributed on the Mel frequency scale. Each filter has the same bandwidth on the Mel frequency scale, and the center frequency of each filter is evenly distributed on the Mel frequency scale. It is then converted back to linear frequency, and the output of the filter group is Where Hm(k) is the frequency response of the mth filter, Sm is the output energy of the mth filter, and m represents the number of any filter;

[0015] The output energy Sm of each filter is taken logarithmically to obtain log(Sm); the filter output after taking the logarithm is subjected to L-point discrete cosine transform, which is expressed by the formula: Get the Mel frequency cepstral coefficient c n , where M is the total number of filters, g is a variable used to index the Mel-frequency cepstral coefficients, and its value range is g=0,1,......L-1;

[0016] For the wavelet packet energy feature, select J decomposition layers and perform J-layer wavelet packet decomposition on the voiceprint signal to obtain 2 J wavelet packet coefficients, obtain the coefficient sequence of the rth wavelet packet node in the fth layer, and calculate it by the formula Get the energy E of the wavelet packet node f,r , where d represents the discrete time index of the node coefficient of the current layer, and Nf,r represents the length of the node coefficient sequence;

[0017] By analogy, the wavelet packet energy of each node in each layer is calculated, and the calculated wavelet packet energy is normalized;

[0018] The Mel frequency cepstral coefficients and wavelet packet energy features are fused to obtain the voiceprint feature vector;

[0019] In this application, the fault classification unit predicts the fault type based on a convolutional neural network, specifically using MobileNetV2 as the fault classification model:

[0020] A large amount of historical voiceprint data with known fault types is extracted from the database as a training set. The fault types corresponding to the historical voiceprint data are annotated. The fault type annotations include metal cutting voiceprints, transformer faults, core grounding faults, circuit breaker faults, mechanical component faults, and cooling system faults. The annotated data is divided into training, validation, and test sets. The cross-entropy loss function is used to measure the difference between the model prediction results and the true labels. The Adam optimizer is used to update the model parameters to minimize the corresponding loss function. The training set is divided into multiple small batches for training, and each small batch contains an equal number of samples. In each training cycle, the model performs forward propagation and backpropagation on all small batches of data to update the model parameters, thereby completing the training and establishment of the fault type model.

[0021] The real-time voiceprint features are input into the trained MobileNetV2 model. The model performs forward propagation calculations based on the input features. Through convolution, pooling, and full connection operations, it outputs the prediction scores for each fault type. The fault type with the highest prediction score is selected as the final classification result, and its classification result is sent to the database and collaborative arbitration module.

[0022] The anti-theft spatiotemporal prediction module extracts features from vibration data and environmental data, and analyzes features in both time and space to identify and determine the threat level of theft. The specific process is as follows:

[0023] Create an LSTM model and output the vibration baseline value. Specifically, the three-axis vibration values are normalized to the range of [0, 1] and then divided into fixed-length time series as input samples for the long-short-term memory network. Historical vibration data in the database is extracted and divided into training and validation sets. The mean square error is selected as the loss function to calculate the square difference between the predicted baseline and the actual value. Iterative training is carried out until the loss converges. The three-axis vibration value at the current time point is input into the LSTM model. Through continuous iteration, the vibration baseline value YF for the next 2 seconds is output. The real-time deviation YS is then calculated according to the absolute difference formula YS = |YT - YF|, where YF represents the real-time three-axis vibration value.

[0024] The propagation path of vibration energy in the box structure is analyzed through graph convolutional networks, and the spatial anomaly index is calculated, specifically:

[0025] The installation position of the sensor on the box is regarded as a node, and (w,z) represents any two node positions. Using the formula Get the conduction path A of vibration energy between nodes w,z , where υw,z Expressed as the structural stiffness between nodes (w,z), ε w,z It is represented as the physical distance between nodes (w, z), σ is the material damping coefficient, and η is the vibration frequency. Similarly, the conduction paths between nodes are obtained, and each conduction path is used as the edge of the graph to construct the graph data.

[0026] The graph data and triaxial vibration values are sent as input to the preset GCN model, and the spatial feature vector of each node is used as output to obtain the feature GCN of each node; the formula is used to calculate The spatial anomaly index SPAI is obtained; where x represents the number of the node feature, B represents the total number of node features, q represents the mean of all node features, and τ represents the standard deviation of all node features;

[0027] The comprehensive threat value TAS is obtained using the formula TAS=0.6×YS+0.4×SPAI, where 0.6 is a fixed constant for the time weight and 0.4 is a fixed constant for the space weight.

[0028] The collaborative arbitration module divides the rule types of box-type substations according to the comprehensive threat value and voiceprint fault type, and processes them. The specific process is as follows:

[0029] The comprehensive threat value is divided into three rule types, including the first rule type, the second rule type and the third rule type. Specifically: if the comprehensive threat value is above 0.7 and the classification result is a metal cutting voiceprint, it is determined to belong to the first rule type, triggering the grid lockout, automatic positioning and alarm; if the comprehensive threat value is within the range of 0.5-0.7, and the classification result is any one of transformer fault, core grounding fault, circuit breaker fault, mechanical component fault and cooling system fault, it is determined to be the second rule type, and a maintenance work order is generated and sent to the corresponding maintenance personnel's mobile terminal; if the comprehensive threat value is below 0.5, there is no voiceprint with a corresponding classification result and the millimeter-wave radar detects that the person's path is less than 2 meters, it is determined to be the third rule type, triggering the camera to take pictures and requesting the maintenance personnel to review.

[0030] Compared with existing technologies, the present invention has the following advantages: the collaborative arbitration module divides rules into three types based on the comprehensive threat value and voiceprint fault type. The system can quickly make corresponding decisions based on different rule types, such as quickly taking measures such as grid lockout and alarm for high-threat theft and generating repair work orders for equipment failures in a timely manner. This targeted processing method greatly improves the system's ability to cope with complex situations. The voiceprint fault diagnosis results and anti-theft monitoring data complement each other, providing a more comprehensive decision-making basis for collaborative arbitration and avoiding decision-making errors caused by data isolation.

[0031] The voiceprint fault diagnosis module comprehensively characterizes voiceprint signals from different dimensions by extracting Mel-frequency cepstral coefficients and wavelet packet energy features, adopts the MobileNetV2 convolutional neural network for fault classification, and uses a large amount of historical voiceprint data to train the model, enabling it to automatically learn the complex mapping relationship between fault type and voiceprint features. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. The following drawings are not intentionally scaled to the actual size, and the focus is on illustrating the main purpose of the present invention.

[0033] Figure 1 This is a module connection diagram of the present invention.

[0034] Figure 2 This is a collaborative arbitration decision logic diagram of the present invention. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts also fall within the scope of protection of the present invention.

[0036] Please refer to Figure 1 As shown, the present invention is a real-time monitoring system for an intelligent box-type substation, which includes a multi-source data acquisition module, a voiceprint fault diagnosis module, an anti-theft spatiotemporal prediction module, a collaborative arbitration module and a database.

[0037] The multi-source data acquisition module collects voiceprint data, vibration data, and environmental data in real time based on multimodal sensors. The specific process is as follows:

[0038] A directional MEMS microphone array is installed on the transformer winding side or the circuit breaker operating cavity to collect the original sound wave signals of equipment operation and abnormal sounds in real time; an ultrasonic sensor is installed near the insulator to collect ultrasonic pulse signals in real time, which can be used to detect microsecond ultrasonic pulse groups caused by insulation defects in the future; three-axis accelerometers are arranged at the four corners of the box or the door axis to collect three-axis vibration values in real time to reflect the intensity and direction of mechanical vibration; a millimeter-wave radar is installed on the rotating pan-tilt head on the top of the box to collect the movement path of the corresponding personnel in real time; the original sound wave signal and ultrasonic pulse signal are integrated into a voiceprint signal, which is sent to the voiceprint fault diagnosis module and voiceprint library and timestamp is added; the three-axis vibration value is sent to the anti-theft spatiotemporal prediction module; the movement path of the personnel is sent to the collaborative arbitration module.

[0039] The voiceprint fault diagnosis module extracts and analyzes the features of the voiceprint data collected during the operation of the box-type substation, accurately identifies the existing fault type, and gives the corresponding confidence level. The specific process is as follows:

[0040] The voiceprint fault diagnosis module is equipped with a pre-processing unit, a feature extraction unit and a fault classification unit;

[0041] The pre-processing unit uses an adaptive filter to separate the noise from the voiceprint signal. The noise includes fan noise and traffic noise outside the substation. The adaptive filter is equipped with an LMS algorithm, whose goal is to minimize the mean square error between the filter output and the expected signal. Specifically:

[0042] Set the input signal to β(n), the expected signal to γ(n), and the output of the adaptive filter to α(n). Use the formula e(n) = γ(n) - α(n) to obtain the error signal e(n). By continuously adjusting the coefficients of the adaptive filter, the mean square value of the error signal is controlled to a minimum, and the noise is filtered out of the signal. The update formula of the filter coefficient is ω(n+1) = ω(n) + ue(n)β(n), where ω(n) represents the filter coefficient and u represents the step factor. The step factor controls the speed and stability of the coefficient update. If the step factor is too large, the adaptive filter converges quickly but causes instability. If the step factor is too small, the adaptive filter converges slowly but has good stability. It should be noted that in the actual operating environment of the box-type substation, the voiceprint signal will be interfered with by various noises. These noises will mask the characteristics of the fault signal and affect the accuracy of fault diagnosis. Therefore, noise reduction processing is required.

[0043] The feature extraction unit extracts the Mel-frequency cepstral coefficients and wavelet packet energy features. Specifically, for the Mel-frequency cepstral coefficients, the input voiceprint signal is pre-emphasized through a first-order FIR filter. The formula is K(n) = β(n)-Q(n-1) to obtain the pre-emphasized signal K(n), where Q is the pre-emphasis coefficient with a fixed value of 0.95. The meanings of the other parameters are the same as those in the adaptive filter formula. The purpose of pre-emphasis is to enhance the energy of the high-frequency part, compensate for the high-frequency attenuation of the signal during transmission, and make the signal spectrum flatter; the pre-emphasized signal The signal is divided into several short frame signals of fixed length, and each short frame signal is subjected to fast Fourier transform FFT to obtain the discrete spectrum X(k) of the short frame signal, where k represents the discrete frequency point after FFT transformation, k=0, 1, ..., N-1, and N-1 represents the total number of discrete frequency points. The discrete spectrum is mapped to the Mel frequency scale, and a group of filters are evenly distributed on the Mel frequency scale. Each filter has the same bandwidth on the Mel frequency scale, and the center frequency of each filter is evenly distributed on the Mel frequency scale. It is then converted back to linear frequency, and the output of the filter group is Where Hm(k) is the frequency response of the mth filter, Sm is the output energy of the mth filter, and m represents the number of any filter;

[0044] The output energy Sm of each filter is taken logarithmically to obtain log(Sm); the filter output after taking the logarithm is subjected to L-point discrete cosine transform. It should be noted that L represents the number of points of discrete cosine transform, and also determines the dimension of the final Mel-frequency cepstral coefficient, which can be expressed as Get the Mel frequency cepstral coefficient c n , where M is the total number of filters, g is a variable used to index the Mel-frequency cepstral coefficient, and its value range is g = 0, 1, ... L-1; it should be noted that the addition of 0.5 in the formula is mainly used to adjust the phase of the cosine function to align it with the center frequency of the Mel filter for more accurate feature extraction;

[0045] For the wavelet packet energy feature, we choose J decomposition layers. It should be noted that the number of decomposition layers determines the fineness of the frequency sub-band division. The higher the number of decomposition layers, the narrower the frequency sub-band, which can provide more refined frequency analysis, but the amount of calculation will also increase. The voiceprint signal is decomposed by J-layer wavelet packets to obtain 2 J wavelet packet coefficients, obtain the coefficient sequence of the rth wavelet packet node in the fth layer, and calculate it by the formula Get the energy E of the wavelet packet node f,r , where d represents the discrete time index of the node coefficient of the current layer, and Nf,r represents the length of the node coefficient sequence;

[0046] By analogy, the wavelet packet energy of each node in each layer is calculated, and the calculated wavelet packet energy is normalized;

[0047] The Mel frequency cepstral coefficients and wavelet packet energy features are fused to obtain the voiceprint feature vector;

[0048] In this application, the fault classification unit predicts the fault type based on a convolutional neural network, specifically using MobileNetV2 as the fault classification model:

[0049] A large amount of historical voiceprint data with known fault types is extracted from the database as a training set. The fault types corresponding to the historical voiceprint data are annotated. The fault type annotations include metal cutting voiceprints, transformer faults, core grounding faults, circuit breaker faults, mechanical component faults, and cooling system faults. The annotated data is divided into training, validation, and test sets. The cross-entropy loss function is used to measure the difference between the model prediction results and the true labels. The Adam optimizer is used to update the model parameters to minimize the corresponding loss function. The training set is divided into multiple small batches for training, and each small batch contains an equal number of samples. In each training cycle, the model performs forward propagation and backpropagation on all small batches of data to update the model parameters, thereby completing the training and establishment of the fault type model.

[0050] The real-time voiceprint features are input into the trained MobileNetV2 model. The model will perform forward propagation calculations based on the input features. Through convolution, pooling, and full connection operations, it outputs the prediction scores for each fault type. The fault type with the highest prediction score is selected as the final classification result, and its classification result is sent to the database and collaborative arbitration module.

[0051] The anti-theft spatiotemporal prediction module extracts features from vibration data and environmental data, and analyzes features in both time and space to identify and determine the threat level of theft. The specific process is as follows:

[0052] Create an LSTM model and output the vibration baseline value. Specifically, the three-axis vibration values are normalized to the range of [0, 1] and then divided into time series of fixed length as input samples for the long short-term memory network. Historical vibration data in the database is extracted and divided into training and validation sets. The mean square error is selected as the loss function to calculate the square difference between the predicted baseline and the actual value. It is trained iteratively until the loss converges. The three-axis vibration value at the current time point is input into the LSTM model. Through continuous iteration, the vibration baseline value YF for the next 2 seconds is output. The real-time deviation YS is then calculated according to the absolute difference formula YS = |YT-YF|, where YF represents the real-time three-axis vibration value. It should be noted that the greater the deviation, the more abnormal the vibration.

[0053] The propagation path of vibration energy in the box structure is analyzed through graph convolutional networks, and the spatial anomaly index is calculated, specifically:

[0054] The installation position of the sensor on the box is regarded as a node, and (w,z) represents any two node positions. Using the formula Get the conduction path A of vibration energy between nodes w,z , where υ w,z Expressed as the structural stiffness between nodes (w,z), ε w,zIt is represented as the physical distance between nodes (w, z), σ is the material damping coefficient, and η is the vibration frequency. Similarly, the conduction paths between nodes are obtained, and each conduction path is used as the edge of the graph to construct the graph data.

[0055] The graph data and triaxial vibration values are sent as input to the preset GCN model, and the spatial feature vector of each node is used as output to obtain the feature GCN of each node; the formula is used to calculate The spatial anomaly index (SPAI) is obtained; where x represents the number of the node feature, B represents the total number of node features, q represents the mean of all node features, and τ represents the standard deviation of all node features. It should be noted that SPAI reflects whether the propagation of vibration energy in the box structure conforms to the normal mode.

[0056] The comprehensive threat value TAS is obtained using the formula TAS=0.6×YS+0.4×SPAI, where 0.6 is a fixed constant for the time weight and 0.4 is a fixed constant for the space weight.

[0057] The collaborative arbitration module divides the rule types of box-type substations according to the comprehensive threat value and voiceprint fault type, and processes them. The specific process is as follows:

[0058] like Figure 2 As shown, the comprehensive threat value is divided into three rule types, including the first rule type, the second rule type and the third rule type. Specifically: if the comprehensive threat value is above 0.7 and the classification result is a metal cutting voiceprint, it is determined to belong to the first rule type, triggering the grid lockout, automatic positioning and alarm; if the comprehensive threat value is within the range of 0.5-0.7, and the classification result is any one of transformer fault, core grounding fault, circuit breaker fault, mechanical component fault and cooling system fault, it is determined to be the second rule type, and a maintenance work order is generated and sent to the corresponding maintenance personnel's mobile terminal; if the comprehensive threat value is below 0.5, there is no voiceprint with the corresponding classification result and the millimeter wave radar detects that the human path is less than 2 meters, it is determined to be the third rule type, triggering the camera to shoot and requesting the maintenance personnel to review.

[0059] The above is an illustration of the present invention and should not be considered as limiting thereof. Although several exemplary embodiments of the present invention have been described, it will be readily understood by those skilled in the art that many modifications may be made to the exemplary embodiments without departing from the novel teachings and advantages of the present invention. Therefore, all such modifications are intended to be included within the scope of the present invention as defined by the claims. It should be understood that the above is an illustration of the present invention and should not be considered as being limited to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present invention is defined by the claims and their equivalents.

Claims

1. A real-time monitoring system for an intelligent box-type substation, comprising a multi-source data acquisition module, a voiceprint fault diagnosis module, an anti-theft spatiotemporal prediction module, a collaborative arbitration module, and a database, characterized by: The voiceprint fault diagnosis module is equipped with a pre-processing unit, a feature extraction unit, and a fault classification unit. It extracts and analyzes the voiceprint data collected during the operation of the box-type substation, and identifies the fault type through the extracted Mel-frequency cepstral coefficients and wavelet packet energy characteristics. The anti-theft spatiotemporal prediction module extracts vibration data features and combines them with feature analysis in time and space dimensions to identify and determine the threat level of theft. The collaborative arbitration module divides the rule types of box-type substations according to the comprehensive threat value and voiceprint fault type, and processes them. Specifically, the comprehensive threat value is divided into three rule types, including the first rule type, the second rule type, and the third rule type. If the comprehensive threat value is above 0.7 and the classification result is a metal cutting voiceprint, it is determined to belong to the first rule type, triggering grid lockout, automatic positioning and alarm; if the comprehensive threat value is within the range of 0.5-0.7 and the classification result is any one of transformer fault, core grounding fault, circuit breaker fault, mechanical component fault and cooling system fault, it is determined to belong to the second rule type, and a maintenance work order is generated and sent to the corresponding maintenance personnel's mobile terminal; If the comprehensive threat value is below 0.5, and there is no corresponding classification result for the voiceprint and millimeter-wave radar that detects a person's path less than 2 meters, the third rule type is determined, and the camera is triggered to take pictures and request maintenance personnel to review.

2. A real-time monitoring system for an intelligent box-type substation according to claim 1, characterized in that: The pre-processing unit processes the voiceprint data collected during the operation of the box-type substation. The specific process is as follows: The noise of the voiceprint signal is separated by an adaptive filter. The noise includes fan noise and traffic noise outside the substation. The adaptive filter is equipped with an LMS algorithm, whose goal is to minimize the mean square error between the filter output and the desired signal. Specifically, the input signal is set to β(n), the desired signal is γ(n), and the output of the adaptive filter is α(n). The error signal e(n) is obtained using the formula e(n) = γ(n) - α(n). By continuously adjusting the coefficients of the adaptive filter, the mean square value of the error signal is controlled to a minimum, and the noise is filtered out from the signal. The update formula of the filter coefficient is ω(n+1)=ω(n)+ue(n)β(n), where ω(n) represents the filter coefficient and u represents the step size factor.

3. A real-time monitoring system for an intelligent box-type substation according to claim 2, characterized in that: The feature extraction unit extracts Mel-frequency cepstral coefficients and wavelet packet energy features, specifically: For the Mel-frequency cepstrum coefficient, the input voiceprint signal is pre-emphasized through a first-order FIR filter. The formula is K(n) = β(n) - Q(n-1) to obtain the pre-emphasized signal K(n), where Q is the pre-emphasis coefficient with a fixed value of 0.

95. The meanings of the other parameters are the same as those of the adaptive filter formula. The pre-emphasized signal is divided into several short frame signals of fixed length. Each short frame signal is subjected to a fast Fourier transform FFT to obtain the discrete spectrum X(k) of the short frame signal. k represents the discrete frequency point after FFT transformation, k = 0, 1, ..., N-1, N-1 represents the total number of discrete frequency points, and the discrete spectrum is mapped to the Mel frequency scale. The center frequency of each filter is evenly distributed on the Mel frequency scale, and then it is converted back to linear frequency. The output of the filter group is Where Hm(k) is the frequency response of the mth filter, Sm is the output energy of the mth filter, and m represents the number of any filter; The output energy Sm of each filter is taken logarithmically to obtain log(Sm); the filter output after taking the logarithm is subjected to L-point discrete cosine transform, which is expressed by the formula: Get the Mel frequency cepstral coefficient c n , where M is the total number of filters, g is a variable used to index the Mel-frequency cepstral coefficients, and its value range is g=0,1,......L-1; For the wavelet packet energy feature, select J decomposition layers and perform J-layer wavelet packet decomposition on the voiceprint signal to obtain 2 J wavelet packet coefficients, obtain the coefficient sequence of the rth wavelet packet node in the fth layer, and calculate it by the formula Get the energy E of the wavelet packet node f,r , where d represents the discrete time index of the node coefficient of the current layer, and Nf,r represents the length of the node coefficient sequence; By analogy, the wavelet packet energy of each node in each layer is calculated, and the calculated wavelet packet energy is normalized; The voiceprint feature vector is obtained by fusing the Mel-frequency cepstral coefficients and wavelet packet energy features.

4. A real-time monitoring system for an intelligent box-type substation according to claim 3, characterized in that: The fault classification unit predicts the fault type based on a convolutional neural network and uses MobileNetV2 as the fault classification model, specifically: A large amount of historical voiceprint data with known fault types is extracted from the database as a training set. The fault types corresponding to the historical voiceprint data are annotated. The fault type annotations include metal cutting voiceprints, transformer faults, core grounding faults, circuit breaker faults, mechanical component faults, and cooling system faults. The annotated data is divided into training, validation, and test sets. The cross-entropy loss function is used to measure the difference between the model prediction results and the true labels. The Adam optimizer is used to update the model parameters to minimize the corresponding loss function. The training set is divided into multiple small batches for training, and each small batch contains an equal number of samples. In each training cycle, the model performs forward propagation and backpropagation on all small batches of data to update the model parameters, thereby completing the training and establishment of the fault type model. The real-time voiceprint features are input into the trained MobileNetV2 model. The model performs forward propagation calculations based on the input features. Through convolution, pooling, and full connection operations, it outputs the prediction scores for each fault type. The fault type with the highest prediction score is selected as the final classification result, and its classification result is sent to the database and collaborative arbitration module.

5. The real-time monitoring system for an intelligent box-type substation according to claim 1 is characterized in that: The anti-theft spatiotemporal prediction module creates an LSTM model and outputs a vibration baseline value. The specific process is as follows: The three-axis vibration values are standardized to the range of [0,1] and then divided into time series of fixed length as input samples of the long short-term memory network. The historical vibration data in the database is extracted and divided into training and validation sets. The mean square error is selected as the loss function to calculate the square difference between the predicted baseline and the actual value. The training is iterated until the loss converges. The three-axis vibration value at the current time point is input into the LSTM model. Through continuous iteration, the vibration baseline value YF within the next 2 seconds is output. The real-time deviation YS is then calculated according to the absolute difference formula YS = |YT-YF|, where YF represents the real-time three-axis vibration value.

6. A real-time monitoring system for an intelligent box-type substation according to claim 5, characterized in that: The anti-theft spatiotemporal prediction module analyzes the propagation path of vibration energy in the box structure through a graph convolutional network and calculates the spatial anomaly index, specifically: The installation position of the sensor on the box is regarded as a node, and (w,z) represents any two node positions. Using the formula Get the conduction path A of vibration energy between nodes w,z , where υ w,z Expressed as the structural stiffness between nodes (w,z), ε w,z It is represented as the physical distance between nodes (w, z), σ is the material damping coefficient, and η is the vibration frequency. Similarly, the conduction paths between nodes are obtained, and each conduction path is used as the edge of the graph to construct the graph data. The graph data and triaxial vibration values are sent as input to the preset GCN model, and the spatial feature vector of each node is used as output to obtain the feature GCN of each node; Calculate using the formula The spatial anomaly index SPAI is obtained; where x represents the number of the node feature, B represents the total number of node features, q represents the mean of all node features, and τ represents the standard deviation of all node features.

7. A real-time monitoring system for an intelligent box-type substation according to claim 6, characterized in that: The anti-theft spatiotemporal prediction module calculates the comprehensive threat value, specifically: The comprehensive threat value TAS is obtained using the formula TAS=0.6×YS+0.4×SPAI, where 0.6 is a fixed constant for the time weight and 0.4 is a fixed constant for the space weight.

8. The real-time monitoring system for an intelligent box-type substation according to claim 1 is characterized in that: The multi-source data acquisition module collects voiceprint data, vibration data, and environmental data in real time based on multimodal sensors. The specific process is as follows: A directional MEMS microphone array is installed on the transformer winding side or the circuit breaker operating cavity to collect the original sound wave signals of equipment operation and abnormal sounds in real time; an ultrasonic sensor is installed near the insulator to collect and ultrasonic pulse signals in real time, which can be used to detect microsecond ultrasonic pulse groups caused by insulation defects in the future; three-axis accelerometers are arranged at the four corners of the box or the door axis to collect three-axis vibration values in real time to reflect the intensity and direction of mechanical vibration; a millimeter-wave radar is installed on the rotating pan-tilt head on the top of the box to collect the movement path of the corresponding personnel in real time; the original sound wave signal and ultrasonic pulse signal are integrated into a voiceprint signal, which is sent to the voiceprint fault diagnosis module and voiceprint library and timestamp is added; the three-axis vibration value is sent to the anti-theft spatiotemporal prediction module; the movement path of the personnel is sent to the collaborative arbitration module.

Images